+ def generate_quizzes(self, nb, model_for_generation, min_ave_seq_logproba):
+ c_quizzes = torch.empty(
+ nb, self.train_w_quizzes.size(1), device=self.device, dtype=torch.int64
+ )
+
+ ar_mask_prompt = torch.zeros(c_quizzes.size(), device=self.device)
+ ar_mask_prompt[:, : ar_mask_prompt.size(1) // 2 + 1] = 1
+ ar_mask_solve = 1 - ar_mask_prompt
+ seq_logproba = torch.empty(ar_mask_prompt.size(0), device=self.device)
+
+ # bracketing of the temperature to get the target logproba if
+ # min_ave_seq_logproba is not None
+
+ temperature = 2
+ d_temperature = 1 / 3
+
+ while True:
+ seq_logproba[...] = 0
+
+ masked_inplace_autoregression(
+ model=model_for_generation,
+ batch_size=self.batch_size,
+ input=c_quizzes,
+ ar_mask=ar_mask_prompt,
+ seq_logproba=seq_logproba,
+ temperature=temperature,
+ deterministic_synthesis=False,
+ # progress_bar_desc="sampling c_quizzes",
+ device=self.device,
+ )
+
+ ave_seq_logproba = seq_logproba.mean()
+
+ masked_inplace_autoregression(
+ model=model_for_generation,
+ batch_size=self.batch_size,
+ input=c_quizzes,
+ ar_mask=ar_mask_solve,
+ seq_logproba=seq_logproba,
+ temperature=temperature,
+ deterministic_synthesis=True,
+ # progress_bar_desc="sampling c_quizzes",
+ device=self.device,
+ )
+
+ # If we do not have target logprobs, get out now
+ if min_ave_seq_logproba is None:
+ break
+
+ # Oh man that's ugly
+ if ave_seq_logproba < min_ave_seq_logproba:
+ if d_temperature > 0:
+ d_temperature *= -1 / 3
+ temperature += d_temperature
+ elif ave_seq_logproba > min_ave_seq_logproba * 0.99:
+ if d_temperature < 0:
+ d_temperature *= -1 / 3
+ temperature += d_temperature
+ else:
+ break
+
+ self.logger(f"changing temperature to {temperature}")
+
+ return c_quizzes, seq_logproba.mean()
+
+ ######################################################################
+
+ def create_c_quizzes(
+ self,
+ nb,
+ model_for_generation,
+ models_for_validation,
+ min_ave_seq_logproba,
+ n_epoch,
+ result_dir,
+ ):
+ c_quizzes, ave_seq_logproba = self.generate_quizzes(
+ nb, model_for_generation, min_ave_seq_logproba
+ )
+
+ nb_correct = self.comput_correctness(c_quizzes, models_for_validation)
+
+ return c_quizzes, nb_correct, ave_seq_logproba
+
+ ######################################################################
+
+ def gang_create_c_quizzes(
+ self,
+ nb,
+ nb_models_for_generation,
+ models,
+ mode,
+ min_ave_seq_logproba,
+ n_epoch,
+ result_dir,
+ ):
+ model_for_generation = Gang(models, nb_models_for_generation, mode)
+ models_for_validation = models
+ return self.create_c_quizzes(
+ nb,
+ model_for_generation,
+ models_for_validation,
+ min_ave_seq_logproba,
+ n_epoch,
+ result_dir,
+ )